Demographic Reporting and the Absence of Hispanic/Latino Representation in Dermatology Artificial Intelligence: A Scoping Review Authors: Sofía Pérez‐Lalinde, Francisco Flores Published: 2026 ABSTRACT ABSTRACTBackgroundArtificial intelligence (AI) is increasingly integrated into dermatology, particularly through image‐based diagnostic tools. However, concerns persist regarding unequal performance across skin tones and limited demographic transparency in model development. Evidence suggests that AI systems trained predominantly on lighter skin types may perform less accurately in darker phototypes, yet reporting of skin tone, race, and ethnicity remains inconsistent across studies.ObjectiveTo map how dermatology AI studies report skin tone, race and ethnicity, Hispanic/Latino identity, and subgroup diagnostic performance.MethodsPubMed, Embase, and Scopus were searched for articles published between January 2020 and November 2025 that applied AI or deep learning to human dermatologic conditions and reported at least one diagnostic performance metric. Two reviewers independently screened studies and extracted data on demographic reporting, dataset composition, and performance across subgroups.ResultsSixteen studies met inclusion criteria. Most datasets were composed predominantly of lighter skin types (Fitzpatrick I–III), with limited inclusion of Fitzpatrick V–VI. Although thirteen studies reported some form of skin‐tone information, reporting approaches varied substantially. Race and ethnicity were inconsistently documented, often using broad categories, and no study explicitly identified Hispanic or Latino participants. Only a minority of studies stratified diagnostic performance by skin tone or demographic subgroups; Among those that performed subgroup analyses, consistent performance disparities were observed, including reduced specificity, increased false‐positive rates, and lower overall diagnostic accuracy in darker skin tones.ConclusionsCurrent dermatology AI literature frequently lacks the demographic transparency necessary to assess algorithmic fairness and generalizability. The underrepresentation of darker skin tones and the absence of Hispanic/Latino identity reporting limit the equitable clinical applicability of existing AI systems. Standardized demographic reporting, intentional dataset diversification, and routine subgroup performance analyses are essential for the development of fair and inclusive dermatology AI tools.Keywords: artificial intelligence, deep learning, dermatology, diagnostic imaging, health equity, healthcare disparities, hispanic, Latino, machine learning, skin pigmentation Introduction Artificial intelligence (AI) has rapidly become a powerful tool in dermatology, with image‐based deep learning (DL), a subset of machine learning that uses multilayered neural networks to automatically learn hierarchical visual features, achieving expert‐level performance in several diagnostic tasks when tested on datasets composed primarily of lighter skin tones [1]. Within this context, convolutional neural networks (CNNs), a class of DL architectures specifically designed for image analysis, have been widely adopted for dermatologic image classification, lesion detection, and disease recognition [2]. More recently, Vision Transformers (ViTs) and hybrid CNN–Transformer architectures have emerged as powerful alternatives for dermatologic image analysis. Unlike CNNs, which incorporate strong inductive biases through local receptive fields and weight sharing, Transformer‐based models rely more heavily on large‐scale training datasets to learn visual representations but have demonstrated state‐of‐the‐art performance across numerous medical image classification tasks [3]. As these architectures become increasingly adopted in dermatology AI, ensuring diverse and representative training datasets becomes even more important to support robust model generalizability across different skin tones and populations. These advances have accelerated interest in AI‐assisted diagnosis and expanded its potential applications across inflammatory dermatoses, benign and malignant lesions, and other common dermatologic conditions. However, as these technologies move closer to routine clinical use, concerns have emerged regarding their generalizability and equitable performance across diverse patient populations. A growing body of evidence demonstrates that CNNs trained predominantly on images from Fitzpatrick skin types I–III (a commonly used clinical classification of skin phototypes based on response to ultraviolet exposure) exhibit significantly lower diagnostic accuracy when applied to darker skin, particularly Fitzpatrick V–VI [2]. These performance disparities have been observed across multiple diagnostic outcomes, including reduced specificity, increased false‐positive rates in darker skin, and overall lower diagnostic accuracy compared with lighter phototypes. Experimental studies further support these findings, showing that classifier performance consistently declines as images originally depicting lighter skin are artificially darkened [4, 5]. These findings support the interpretation that insufficient representation of darker skin tones in training datasets is a key driver of diagnostic inequity. Moreover, recent analyses reveal that racial and ethnic data are rarely reported in this body of literature, with Hispanic and Latino populations notably underrepresented despite their substantial burden of dermatologic conditions and diagnostic delay [6]. This lack of demographic transparency further limits the ability to evaluate algorithmic fairness across clinically relevant subgroups. Beyond performance disparities, inconsistent demographic reporting represents a methodological limitation that restricts cross‐study comparison and obscures systematic sources of bias. Beyond differences in dataset representation, dermatologic diseases frequently exhibit distinct clinical manifestations across skin tones. In skin of color, erythema is often less conspicuous, inflammatory disorders may present with greater pigmentary alteration, and certain conditions such as keloids and specific hyperkeratotic phenotypes occur more frequently or with different clinical characteristics [7]. Likewise, dermoscopic features of skin cancers may vary according to pigmentation and population‐specific characteristics, potentially affecting both clinician interpretation and AI‐based image classification [8]. These variations underscore the importance of developing and validating AI systems using diverse datasets that adequately capture the full spectrum of dermatologic presentations across skin tones. Considering the rapid integration of AI into dermatologic practice and the unequal benefits observed across skin tones, there is a critical need to systematically assess how current studies report skin tone, racial and ethnic representation, and subgroup performance metrics. Accordingly, this scoping review seeks to map how studies using human dermatologic data report skin tone, racial and ethnic representation, and subgroup diagnostic performance across the dermatology AI literature, in order to identify gaps and inform more transparent and equitable AI development. Importantly, this review focuses not only on reported diagnostic performance, but also on how demographic variables are documented, categorized, or omitted across studies. Material and Methods This scoping review was conducted following the Joanna Briggs Institute (JBI) methodology and is reported according to the PRISMA extension for scoping reviews (PRISMA‐ScR) guidelines [9, 10]. The review question and eligibility criteria were structured using the Population‐Concept‐Context (PCC) framework. The population included studies involving human participants with dermatologic conditions or real‐world dermatologic image datasets. The concept focused on the reporting of skin tone, race, ethnicity, Hispanic/Latino identity, and subgroup diagnostic performance. The context encompassed all dermatology AI settings, without restrictions on clinical environment or geographic location. A comprehensive search was conducted in PubMed, Embase, and Scopus for articles published between January 1, 2020, and November 30, 2025, and therefore included all records indexed in the selected databases up to that date. Studies published or indexed after completion of the search were not eligible for inclusion. PubMed, Embase, and Scopus were selected because they collectively provide broad coverage of the biomedical, clinical, and interdisciplinary literature relevant to dermatology and AI. These databases have been widely used in evidence syntheses involving medical AI and were considered sufficient to capture the majority of peer‐reviewed studies addressing the research question. Although additional databases such as Web of Science may have identified a limited number of additional records, the selected databases were deemed appropriate for the objectives of this scoping review. The search was conducted using terms related to AI, DL, and dermatology. Terms related to skin tone or racial and ethnic categories were intentionally excluded to avoid filtering out studies that did not report these variables. Studies were eligible if they were original, peer‐reviewed articles evaluating AI‐based diagnostic models in dermatology using human image datasets and reporting diagnostic performance outcomes. Demographic reporting (e.g., skin phototype, race, ethnicity, or geographic origin) was not required for study inclusion because assessment of the completeness of demographic reporting constituted one of the primary objectives of this scoping review. The detailed eligibility criteria are summarized in Table 1. Reviews, commentaries, conference abstracts, dataset‐only papers, animal studies, studies involving non‐human subjects, and articles using non‐clinical or synthetic dermatologic data (e.g., simulated images or non–patient‐derived datasets) were excluded. The detailed search strategies for all databases are presented in Appendix I. TABLE 1Eligibility criteria.Inclusion criteriaExclusion criteriaOriginal, peer‐reviewed research articles.Reviews, systematic reviews, meta‐analyses, editorials, letters, conference abstracts, case reports, and book chapters.Studies evaluating AI, DL, CNNs, ViTs, or other machine learning models for dermatologic diagnosis or classification.Studies unrelated to dermatology or not evaluating AI‐based diagnostic models.Studies using human dermatologic image datasets (clinical photographs, dermoscopic images, or other dermatologic imaging modalities).Studies using exclusively synthetic, simulated, animal, or non‐human datasets.Studies reporting diagnostic performance of AI models.Studies without sufficient methodological information or without evaluation of model performance.Articles published in English or Spanish between January 2020 and November 30, 2025.Articles published in languages other than English and Spanish, or outside the predefined search period. Eligibility criteria. All records were imported into Rayyan, a web‐based systematic review screening platform designed to support blinded and independent screening by multiple reviewers. Rayyan was used for duplicate removal, screening organization, and identification of conflicts between reviewers. Study selection was performed independently by two reviewers, who screened all titles, abstracts, and subsequently the full‐text articles according to the predefined eligibility criteria. Any discrepancies regarding study eligibility were resolved through discussion and consensus. No AI tools were used to screen, evaluate, or select studies. Data extraction was performed using a structured charting form capturing study characteristics, dermatologic condition, country of dataset origin, reporting of skin tone, race and ethnicity, Hispanic/Latino identity, and any statements related to dataset diversity, bias, or limitations. Extracted data were summarized descriptively and no formal risk‐of‐bias assessment was performed in accordance with scoping review methodology. Results Search Results The database search initially identified 550 records, of which 508 remained after deduplication. Following title and abstract screening, 122 articles were selected for full‐text review. Of these, 16 studies met the inclusion criteria and were included in the final synthesis. Studies were excluded if they focused on non‐human subjects (e.g., animal or in vitro dermatology models) or if they used non‐clinical dermatologic data, such as synthetic images, simulations, or non–patient‐derived datasets. The study selection process is presented in the PRISMA‐ScR flow diagram (Figure 1). PRISMA‐ScR flow diagram of study selection. Characteristics of Included Sources of Evidence The 16 included studies were published between 2020 and 2025 and covered a diverse range of dermatologic conditions, including inflammatory dermatoses, benign and malignant lesions, melanocytic lesions, acne, mpox, and general dermatologic triage tasks. Studies originated from a broad geographic range, most commonly China, the United States, Switzerland, and multinational collaborations. All included studies evaluated DL–based image classification models, including CNNs, ViTs, or hybrid architectures, using human dermatologic image datasets such as clinical photographs, dermoscopic images, or publicly available image repositories. Detailed characteristics of each study, including dermatologic focus, model performance, reporting of skin tone and race/ethnicity, and statements related to bias or dataset limitations, are summarized in Table 2. TABLE 2Summary of included studies.Author, year, countryDermatologic conditionMain findings and AI applicationsSkin tone data reportedLimitations and biasMevorach et al., 2025 – Italy [11]Various dermatologic conditionsThe AI model achieved the highest overall diagnostic performance, with superior specificity compared with expert and non‐expert clinicians, who showed higher sensitivity but more false positives. Performance was consistently lower for actinic keratosis and benign keratosis‐like lesions.Fitzpatrick I–III predominantly.Limited skin‐type diversity (I–III), increasing the risk of algorithmic bias and reduced generalizability.Chong et al., 2022 – China [12]Inflammatory dermatoses (psoriasis vs seborrheic dermatitis)The DL model outperformed five dermatologists in sensitivity, specificity, and AUC for differentiating scalp psoriasis from seborrheic dermatitis, and significantly improved the diagnostic performance of non‐expert clinicians.Fair‐skinned Chinese population; Fitzpatrick types not explicitly reported. Single‐country dataset composed exclusively of fair‐skinned patients, limiting generalizability across skin tones and populations. Wu et al., 2020 – China [13]Inflammatory dermatoses (psoriasis vs eczema)The DL model achieved very high diagnostic performance for differentiating psoriasis, eczema, and healthy skin, clearly outperforming alternative DL architectures.Chinese population; Fitzpatrick skin types not explicitly reported.Thieme et al., 2023 – USA [14]Mpox vs non‐mpox skin lesionsThe CNN achieved high diagnostic performance for differentiating mpox from non‐mpox lesions and maintained robust accuracy across datasets and body regions.Fitzpatrick I–VI; reduced specificity and higher false‐positive rates in Fitzpatrick V–VI.Decreased specificity in darker skin tones indicates skin‐tone–related performance disparities.Huang et al., 2023 – China [15]PsoriasisThe deep learning model outperformed experienced dermatologists in predicting PASI scores and severity trends, achieving high trend‐prediction accuracy in longitudinal severity assessment.Not reported; authors acknowledge potential influence of skin color.Clinical data set was limited to patients of Chinese origin, requiring retraining for other populations.Le Lay et al., 2024 ‐ France [16] Non‐cancerous dermatoses The AI system demonstrated improved diagnostic accuracy over earlier versions, reaching a sensitivity of 88% and specificity of 90, supporting its potential as a reliable tool for triage and diagnosis of common skin conditions. Fitzpatrick II–V; Fitzpatrick I and VI excluded. Recruitment bias due to exclusion of Fitzpatrick I and VI, limiting generalizability across the full skin tone spectrum.Oloruntoba et al., 2025 – Denmark [17]Benign and malignant lesionsBoth clinicians and the CNN achieved their highest diagnostic performance when using high‐quality dermoscopic images, while enhanced or HDR‐style images consistently reduced performance.Predominantly Fitzpatrick II–III.Dataset derived exclusively from Danish patients with limited skin tone and lesion diversity, restricting generalizability.Han et al., 2020 – Multicenter [18]Various dermatologic conditionsThe AI system demonstrated strong performance in malignancy prediction, disease classification, and treatment recommendation, and significantly improved diagnostic accuracy for clinicians and non‐expert users.Predominantly Asian and Caucasian populations; darker phototypes not represented. Fitzpatrick skin types not explicitly reported.Limited representation of darker skin tones reduces generalizability and may introduce population‐specific bias.Wang et al., 2020 – China [19]Various dermatologic conditionshe deep learning classifier achieved diagnostic performance comparable to 164 dermatologists across multiple lesion types in both multiclass and binary classification tasks.Fitzpatrick IV only (Asian population).Single‐center dataset restricted to a single skin type, limiting external validity across other phototypes.Rikhye et al., 2025 – USA [20]Various dermatologic conditionsThe AI model showed comparable performance on clinician‐taken and patient‐taken images, with accuracy improving after resampling datasets to match the training distribution.Representative of the San Francisco Bay Area population; Fitzpatrick skin types not explicitly reported.Validation on broader and more diverse populations is needed to assess generalizability across skin tones and settings.Goessinger et al., 2024 – Switzerland [21]Melanocytic lesionsCommercial CNNs showed moderate‐to‐good reliability across repeated dermoscopic images; averaging multiple captures improved sensitivity and overall robustness.Fitzpatrick I–III; Fitzpatrick >III not included.Restricted applicability to darker phototypes due to limited skin tone diversity.Li et al., 2023 – China [22]Acne vulgarisThe AI system accurately detected multiple acne lesion types and demonstrated strong agreement with clinician‐assessed severity, with improved performance when lesion identification was integrated into grading.Fitzpatrick II‐IV.Single‐population dataset with limited skin tone range and imbalance in acne severity categories.Amruthalingam et al., 2022 – Switzerland [23]Various dermatologic conditionsAutomated anatomical region mapping improved diagnostic precision and sensitivity when integrated into AI‐based skin image classification models.Fitzpatrick I‐III.Limited skin tone diversity prevented assessment of performance across darker phototypes.Amin et al., 2025 – Multicenter [24]Skin lesionsA boundary‐aware segmentation model combined with a transformer‐based classifier achieved very high segmentation performance and classification accuracy (> 98%) across multiple public datasets, supporting improved lesion analysis through integration of local and global features.Higher diagnostic accuracy on light skin. Fitzpatrick skin types not explicitly reported.Neural networks replicate biases present in training data, resulting in reduced performance on darker skin tones.Shen et al., 2024 – China [25]Various dermatologic conditionsA contrastive‐learning model trained on large‐scale internet images outperformed baseline methods and demonstrated strong transferability to emerging diseases, including effective monkeypox detection with minimal additional labeled data.Higher performance observed among East Asian populations. Fitzpatrick skin types not explicitly reported.Limited skin‐tone diversity and restricted applicability to darker phototypes.Kamulegeya et al., 2023 – Uganda [26]Various dermatologic conditionsThe AI system showed very low diagnostic accuracy on Fitzpatrick VI skin compared with substantially better performance on lighter phototypes, revealing large performance disparities across conditions.Fitzpatrick VI.Study explicitly designed as testing AI bias on Black skin. Note: None of the included studies explicitly reported Hispanic or Latino identity or provided subgroup analyses for Hispanic/Latino participants. Summary of included studies. Single‐country dataset composed exclusively of fair‐skinned patients, limiting generalizability across skin tones and populations. Non‐cancerous dermatoses Fitzpatrick II–V; Fitzpatrick I and VI excluded. Note: None of the included studies explicitly reported Hispanic or Latino identity or provided subgroup analyses for Hispanic/Latino participants. Reporting of Skin Tone Most studies (n = 13) reported some form of skin tone information; however, the depth, methodology, and consistency of reporting varied substantially. The majority of datasets were composed predominantly of lighter skin types, particularly Fitzpatrick I–III. Only a small subset of studies included Fitzpatrick V–VI skin, and in many cases representation was minimal or restricted to validation subsets. Several studies explicitly noted lower diagnostic performance on darker skin or acknowledged insufficient representation of darker phototypes within their datasets. Reporting of Race and Ethnicity Reporting of race and ethnicity was inconsistent and generally limited. Most studies described their study populations using broad categories such as “Chinese,” “Asian,” “Caucasian,” or “local population,” without providing detailed racial or ethnic breakdowns. Only one study explicitly included Black African patients (Fitzpatrick VI), and none of the included studies explicitly reported Hispanic or Latino identity. Consequently, the representation of Hispanic/Latino populations could not be assessed. Subgroup Performance Metrics Only a minority of studies stratified AI performance by skin tone or racial and ethnic subgroups. Among those that conducted subgroup analyses, consistent performance disparities were observed, including reduced specificity or increased false‐positive rates in darker skin, lower diagnostic accuracy in Fitzpatrick VI skin compared with lighter phototypes, and performance differences aligned with population demographics, such as higher accuracy in East Asian populations. Among the limited number of studies that reported subgroup analyses, diagnostic performance generally favored lighter skin types, although the number of studies was too small to draw definitive conclusions regarding the magnitude or consistency of these disparities. Geographic Patterns and Dataset Composition Most datasets were derived from single‐country sources with relatively homogeneous populations, such as China, Switzerland, or Denmark. Only a small number of studies incorporated multicenter or multi‐ethnic datasets. Limited geographic and demographic diversity was frequently cited as a constraint on model generalizability and fairness. Acknowledgment of Bias and Dataset Limitations Most included studies explicitly acknowledged concerns related to bias and limited dataset diversity. Authors frequently reported that their datasets were dominated by lighter skin tones (particularly Fitzpatrick I–III), with minimal or no representation of darker phototypes (Fitzpatrick >IV). Several studies noted reduced diagnostic performance on darker skin and emphasized the need for broader, more representative datasets to improve fairness and generalizability. Collectively, these findings highlight persistent gaps in demographic representation and limited evaluation of algorithmic fairness in current dermatology AI research. Discussion This scoping review highlights persistent and clinically relevant gaps in demographic representation and reporting within dermatology AI research, despite rapid advances in algorithmic performance and growing enthusiasm for AI‐assisted diagnostic tools. While the field has made substantial methodological progress, our findings suggest that equity considerations, particularly related to skin tone and demographic diversity, remain insufficiently integrated into model development, validation, and reporting practices. Across the included studies, datasets were overwhelmingly composed of lighter skin tones, with limited inclusion of darker phototypes and minimal stratification of diagnostic performance across demographic subgroups. Most notably, none of the reviewed studies explicitly reported Hispanic or Latino identity, revealing a critical and underexplored blind spot in current dermatology AI literature. Together, these findings raise important concerns regarding the generalizability, fairness, and clinical safety of existing AI systems. Skin Tone Representation and Algorithmic Performance Our results are consistent with a growing body of evidence suggesting that dermatology AI systems trained predominantly on lighter skin tones may exhibit reduced diagnostic performance when applied to darker phototypes. Among the minority of studies that conducted formal subgroup analyses, reported findings suggested lower specificity, increased false‐positive rates, and reduced overall accuracy in darker skin types. However, because only a limited number of studies performed these analyses, the available evidence remains insufficient to determine the magnitude and consistency of these differences across all dermatology AI applications. These findings align with prior experimental and computational studies showing that CNNs are highly sensitive to visual features correlated with pigmentation and contrast, making them particularly vulnerable to bias when training data lack sufficient diversity. Importantly, several studies included in this review explicitly acknowledged limited skin tone diversity as a key limitation, suggesting increasing awareness within the field. However, acknowledgment alone does not mitigate bias. Without intentional dataset diversification and standardized reporting of skin tone distribution, such limitations persist across successive model iterations. As a result, AI tools risk reinforcing existing diagnostic inequities. This issue is particularly relevant in skin of color dermatology, where variations in erythema, pigmentation patterns, dermoscopic features, and disease phenotype may alter the visual characteristics on which AI algorithms rely [7]. An additional methodological challenge is the subjective assessment of skin tone across dermatology AI studies. Many studies rely on the Fitzpatrick skin phototype classification, which was originally developed to describe cutaneous response to ultraviolet exposure rather than to objectively quantify skin pigmentation [27]. Consequently, skin tone classification may vary between observers or be inferred from clinical images without standardized assessment, introducing potential misclassification and reducing comparability across datasets. These inconsistencies may further complicate the evaluation of algorithmic fairness and highlight the need for standardized and reproducible approaches to skin tone annotation in future AI research. Race, Ethnicity, and the Absence of Hispanic/Latino Identity Reporting Beyond skin tone, this review reveals substantial inconsistency and imprecision in the reporting of race and ethnicity. When demographic data were reported, categories were often broad (e.g., “Asian,” “Caucasian,” or “local population”) and poorly defined, limiting interpretability and cross‐study comparisons. Critically, none of the included studies explicitly reported Hispanic or Latino identity as a distinct demographic variable. Consequently, it is not possible to determine whether Hispanic or Latino participants were included but not identified, or whether they were absent from the study populations altogether. This lack of demographic reporting limits the ability to assess algorithmic performance and equity in this population. This absence is particularly concerning given that Hispanic and Latino identity does not map directly onto skin tone or Fitzpatrick classification. Individuals who identify as Hispanic or Latino may span the full spectrum of skin pigmentation and may experience distinct structural, socioeconomic, linguistic, and healthcare access barriers that influence dermatologic outcomes independently of skin tone. The failure to capture this identity therefore represents a missed opportunity to assess AI performance in a population that is both epidemiologically significant and historically underserved in dermatologic care. Moreover, the lack of Hispanic/Latino reporting reflects broader challenges in global AI research, including reliance on geographically narrow datasets, limited engagement with sociocultural determinants of health, and inconsistent application of demographic standards. Addressing these gaps will be essential for developing AI systems that are both clinically valid and socially responsible. Implications for Fairness, Generalizability, and Clinical Deployment The absence of standardized demographic reporting substantially limits the ability to evaluate algorithmic fairness and real‐world applicability. AI systems validated primarily on homogeneous datasets may demonstrate high aggregate accuracy while masking clinically meaningful performance disparities in underrepresented populations. In dermatology, where visual diagnosis is central and disease presentation varies by pigmentation and context, such disparities carry tangible clinical risks. Importantly, challenges related to generalizability are not unique to AI systems. Dermatologists' diagnostic reasoning may also be influenced by the country and healthcare setting in which they practice, reflecting differences in patient populations, disease prevalence, cultural norms, and healthcare systems [28]. Consequently, both clinicians and AI models may experience reduced diagnostic performance when applied to populations that differ substantially from those on which their expertise or training was developed. This parallel further highlights the importance of geographically and demographically diverse datasets for both clinical education and AI development. From a regulatory and ethical standpoint, these findings underscore the need for clearer reporting standards and accountability mechanisms. Clinicians, patients, and regulatory bodies require transparent information regarding dataset composition and subgroup performance to make informed decisions about clinical deployment. Without routine subgroup analyses and explicit documentation of demographic characteristics, it remains difficult to determine whether AI systems can be safely and equitably integrated into everyday dermatologic practice. Future Directions Future dermatology AI research should prioritize standardized, transparent reporting of skin tone, race, and ethnicity, with explicit inclusion of Hispanic/Latino identity where relevant. Prospective, multicenter datasets that intentionally recruit participants across the full range of Fitzpatrick phototypes and diverse demographic backgrounds are needed to support equitable model development and validation. In addition, routine reporting of subgroup performance metrics should be incorporated into publication standards and journal requirements, enabling meaningful assessment of algorithmic fairness and facilitating more responsible clinical translation. Strengths and Limitations This review has several strengths. To our knowledge, it is the first scoping review specifically examining demographic reporting, Hispanic/Latino representation, and subgroup performance within dermatology AI literature using a structured PRISMA‐ScR methodology. This review has several limitations that should be considered when interpreting the findings. First, although the search strategy was designed to maximize sensitivity within three major biomedical databases (PubMed, Embase, and Scopus), it may not have captured all relevant studies indexed elsewhere or published outside the predefined search period. Consequently, only sixteen studies met the eligibility criteria, reflecting both the relatively recent development of dermatology AI research and the strict inclusion criteria requiring human clinical data and reported diagnostic performance metrics. Second, as a scoping review, this study did not perform a formal risk‐of‐bias assessment or quantitative synthesis. Finally, the conclusions are inherently dependent on the quality and completeness of demographic reporting in the included studies. In addition, the search was limited to three major bibliographic databases. Although these databases provide extensive coverage of biomedical research, relevant studies indexed exclusively in other databases, such as Web of Science, may not have been identified. Conclusion This scoping review identifies substantial and persistent gaps in the reporting of skin tone, race, ethnicity, and Hispanic/Latino identity within dermatology AI research, alongside limited evaluation of diagnostic performance across diverse demographic subgroups. Collectively, these deficiencies indicate that the current evidence base lacks the demographic transparency necessary to rigorously assess algorithmic fairness and to support equitable clinical implementation. Without more representative datasets and standardized reporting practices, the clinical applicability of dermatology AI systems across diverse populations remains uncertain. Future research should prioritize standardized and transparent reporting of demographic variables, including skin tone and detailed racial and ethnic categories, to enable meaningful evaluation of model performance across populations. Studies that intentionally recruit participants spanning the full spectrum of Fitzpatrick phototypes (and that explicitly report Hispanic/Latino identity and, where appropriate, ensure adequate representation of Hispanic/Latino populations, whose inclusion cannot currently be assessed because demographic reporting is largely absent) are critically needed. The development of prospective, multicenter datasets with harmonized skin tone annotation frameworks would strengthen the generalizability of dermatology AI models and support future systematic evaluations grounded in more robust and inclusive evidence. Additionally, methodological transparency, particularly with respect to subgroup analyses, should be integrated into reporting guidelines and journal standards to promote responsible and equitable clinical translation. Funding The authors declare no funding. Conflicts of Interest The authors declare no conflicts of interest. Acknowledgments The authors thank the faculty and research mentors at the University of Miami for their academic guidance. We also acknowledge the support of colleagues who provided feedback on the conceptualization and early drafts of this review. Search strategy DatabaseSearch equationResultsPubMed (“deep learning”[ti] OR “artificial intelligence”[ti]) AND (dermatology[tiab] OR dermatologic*[tiab]) AND (“2020/01/01”[PDAT]: “2025/11/30”[PDAT]) AND (english[lang] OR spanish[lang]) 240Embase (“deep learning”/exp OR “deep learning”:ti, ab) AND (“dermatology”/exp OR dermatology:ti, ab) AND [2020‐2025]/py AND ([english]/lim OR [spanish]/lim) AND [article]/lim 140Scopus (“dermatology” MeSH Terms) AND (“artificial intelligence” MeSH Terms) AND PUBYEAR > 2020 AND PUBYEAR < 2026 170 (“deep learning”[ti] OR “artificial intelligence”[ti]) AND (dermatology[tiab] OR dermatologic*[tiab]) AND (“2020/01/01”[PDAT]: “2025/11/30”[PDAT]) AND (english[lang] OR spanish[lang]) (“deep learning”/exp OR “deep learning”:ti, ab) AND (“dermatology”/exp OR dermatology:ti, ab) AND [2020‐2025]/py AND ([english]/lim OR [spanish]/lim) AND [article]/lim (“dermatology” MeSH Terms) AND (“artificial intelligence” MeSH Terms) AND PUBYEAR > 2020 AND PUBYEAR < 2026 Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request. References Associated Data Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.